Research articles

Despite their efficiency advantages, the performance of photonic neural networks is hampered by the accumulation of inherent systematic errors. Zheng et al. propose a dual backpropagation training approach, which allows the network to adapt to systematic errors, thus outperforming state-of-the-art in situ training approaches.

Local methods of explainable artificial intelligence identify where important features or inputs occur, while global methods try to understand what features or concepts have been learned by a model. The authors propose a concept-level explanation method that bridges the local and global perspectives, enabling more comprehensive and human-understandable explanations.

With the advances in neural language models, the question arises if some models align better with human processing than others. Golan et al. identify sentences that language models disagree about and use them to compare the shortcomings of different language models.

An outstanding challenge in materials science is doing large-scale simulations with complex electron interactions. Deng and colleagues introduce a universal graph neural network-based interatomic potential integrating atomic magnetic moments as charge constraints, which allows for capturing subtle chemical properties in several lithium-based solid-state materials

Generating novel molecules that bind to specific protein targets is a challenging but important task in computational drug design. Zhang and colleagues present a molecular generation method based on hierarchical auto-regression.

For virtual protein docking, an accurate scoring function is necessary that evaluates how likely a protein conformation is. Stebliankin and colleagues present a method based on vision transformers that provides a more accurate score by evaluating individual binding interfaces as multi-channel images.

Achieving sequential robotic actions involving different manipulation skills is an open challenge that is critical to enable robots to interact meaningfully with their physical environment. Triantafyllidis and colleagues present a hierarchical learning framework based on an ensemble of specialized neural networks to solve complex long-horizon manipulation tasks.

Traditional feedback-state selection in robot learning is empirical and requires substantial engineering efforts. Yu et al. develop a quantitative and systematic state-importance analysis, revealing crucial feedback signals for learning locomotion skills.

Tandem mass spectroscopy is a useful tool to identify metabolites but is limited by the capability of computational methods to annotate peaks with chemical structures when spectra are dissimilar to previously observed spectra. Goldman and colleagues use a transformer-based method to annotate chemical structure fragments, thereby incorporating domain insights into its architecture, and to simultaneously predict the structure of the metabolite and its fragments from the spectrum.

The heterogeneous and compartmentalized environments within living cells make it difficult to deploy theranostic agents with precise spatiotemporal accuracy. Zhao et al. demonstrate a DNA framework state machine that can switch among multiple structural states according to the temporal sequence of molecular cues, enabling temporally controlled CRISPR–Cas9 targeting in living mammalian cells.

A challenging problem in deep learning consists in developing theoretical frameworks suitable to study generalization. Feng and colleagues uncover a duality relation between neuron activities and weights in deep learning neural networks, and use it to show that sharpness of the loss landscape and norm of the solution act together in determining its generalization performance.

Deep learning applied to live-cell images of patient-derived neurons aids predicting underlying mechanisms and gains insights into neurodegenerative diseases, facilitating the understanding of mechanistic heterogeneity. D’Sa and colleagues use patient-derived stem cell models, high-throughput imaging and machine learning algorithms to investigate Parkinson’s disease subtyping.

Microscopic imaging and holography aim to decrease reliance on labelled experimental training data, which can introduce biases, be time-consuming and costly to prepare, and may lack real-world diversity. Huang et al. develop a physics-driven self-supervised model that eliminates the need for labelled or experimental training data, demonstrating superior generalization on the reconstruction of experimental holograms of various samples.

The tendency of machine learning algorithms to learn biases from training data calls for methods to mitigate unfairness before deployment to healthcare and other applications. Yang et al. propose a reinforcement-learning-based method for algorithmic bias mitigation and demonstrate it on COVID-19 screening and patient discharge prediction tasks.

Algorithmic super-resolution in the context of fluorescence microscopy is challenging due to the difficulty to reliably represent biological nanostructures in synthetically generated images. Bouchard and colleagues propose a deep learning model for live-cell imaging that can leverage auxiliary microscopy imaging tasks to guide and enhance reconstruction, while preserving the biological features of interest.

To ensure that a machine learning model has learned the intended features, it can be useful to have an explanation of why a specific output was given. Slack et al. have created a conversational environment, based on language models and feature importance, which can interactively explore explanations with questions asked in natural language.

Optimal control of quantum many-body systems is needed to make use of quantum technologies, but is challenging due to the exponentially large dimension of the Hilbert space as a function of the number of qubits. Metz and Bukov propose a framework combining matrix product states and reinforcement learning that allows control of a larger number of interacting quantum particles than achievable with standard neural-network-based methods.

There are currently promising developments in deep learning for protein design, with applications in drug discovery and synthetic biology. For more efficient exploration of the design space, Wang et al. demonstrate a reinforcement learning method, EvoZero, for directed evolution in protein engineering towards desired functional or structure-related properties.

Out of the large number of neoepitopes, few elicit an immune response from the major histocompatibility complex. To predict which neoepitopes can be effective, Albert and colleagues present a method based on long short-term memory ensembles and transfer learning from immunogenicity assays.

Integrating gene expression across tissues is crucial for understanding coordinated biological mechanisms. Viñas et al. present a neural network for multi-tissue imputation of gene expression, exploiting the shared regulatory architecture of tissues.